+
+ | Project |
+ Hermes Agent (Nous Research) |
+
+
+ | Integration Type |
+ Frontend UI / Terminal Agent Harness |
+
+
+ | Connection Method |
+ Custom provider, OpenAI-compatible (api_mode: chat_completions) |
+
+
+ | Olla Endpoint |
+ http://localhost:40114/olla/openai/v1 |
+
+
+ | Sticky Sessions |
+ Needs a community plugin to send X-Olla-Session-ID (see below) |
+
+
+
+## Why this page exists
+
+Olla's [sticky sessions](../../concepts/sticky-sessions.md) pin the turns of a conversation to the same backend so the backend can reuse its KV-cache instead of re-ingesting the whole prompt each turn. The most reliable key source is `session_header`, which reads the `X-Olla-Session-ID` request header.
+
+Hermes can already attach custom headers to a custom provider. Both `model.default_headers` (every request) and a per-provider `extra_headers` map (scoped to one endpoint) are supported natively:
+
+```yaml
+model:
+ providers:
+ olla:
+ base_url: http://localhost:40114/olla/openai/v1
+ extra_headers:
+ X-Client-Name: hermes-agent
+```
+
+Those values are **static**, which is the catch. A fixed `X-Olla-Session-ID` would give every conversation the same key, pinning all of your traffic to a single backend permanently. That is worse than leaving the header off, where Olla at least falls back to `prefix_hash`.
+
+What sticky sessions need is a value that changes per conversation, and Hermes has no built-in way to set one. Filling that gap needs plugin middleware.
+
+This recipe was raised in [issue #195](https://github.com/thushan/olla/issues/195), by a user running Olla as the router for a local inference cluster.
+
+## The community plugin
+
+!!! warning "Not maintained by Olla"
+ [`shared-goals/hermes-custom-header-plugin`](https://github.com/shared-goals/hermes-custom-header-plugin) is a third-party plugin. It is not written, maintained, reviewed or supported by the Olla project. Read the source before installing it, and raise plugin problems on the plugin's own issue tracker rather than ours.
+
+ Olla implements nothing specific to Hermes. Anything that can send `X-Olla-Session-ID` gets the same behaviour. If Hermes gains native custom-header support, you can opt for that (and we'll update the docs too!).
+
+The plugin registers [`llm_request` middleware](https://github.com/NousResearch/hermes-agent/blob/main/docs/middleware/README.md), a documented Hermes extension point that runs just before each provider call and can rewrite the request. Middleware callbacks receive runtime context including `session_id` and `model`, which is what makes a per-conversation value possible at all.
+
+For provider identities you explicitly name, the plugin adds one header whose value is the current Hermes session ID and the effective model joined with a colon:
+
+```text
+20260715_210001_a1b2c3:coder
+```
+
+Olla hashes whatever arrives in that header with 64-bit FNV-1a, so the exact format does not matter to Olla. What matters is that the value is stable across the turns of one conversation and differs between conversations, which it is.
+
+Note that Olla already scopes session keys by model internally, so the `: